CASE STUDY / COMPUTER VISIONAPRIL 2026

Geometry-aware risk estimation for vehicle interactions.

A four-stage vision pipeline showing why perspective correction matters when risk is inferred from a single traffic camera.

MY ROLE

System integration · trajectory features · risk formulation · ablations · visualization · report writing

Side-by-side traffic frame comparing image-plane risk with geometry-aware risk

1,090video frames

31,158vehicle detections

0.114pair-set Jaccard

8×fewer unique pairs

01 / THE QUESTION

Pixel distance is not physical distance.

Traffic cameras flatten a three-dimensional scene. Near the horizon, vehicles that are far apart in the world can sit only a few pixels apart; near the camera, equivalent motion spans far more pixels. A risk model operating directly in image space inherits that distortion.

The project isolates this failure mode by comparing two risk scorers with the same pairwise features—distance, closing speed, and time-to-collision—but different coordinate systems.

02 / THE PIPELINE

Detect

YOLOv8n identifies cars, buses, and trucks.

→

Track

ByteTrack maintains identities and trajectories.

→

Project

A normalized DLT homography maps motion to BEV.

→

Score

Distance, closing speed, and TTC flag risk.

Traffic frame with four ground-plane calibration points — Four road-plane correspondences anchor the camera-to-ground homography.

Vehicle trajectories projected into bird’s-eye-view coordinates — Projected trajectories form parallel road streams inside the calibrated region.

03 / WHAT CHANGED

Geometry removed an order of magnitude of noise.

The image-plane baseline flagged 5,474 risky pair-frames across 830 unique pairs. The geometry-aware variant flagged 610 pair-frames across 108 unique pairs. The two methods agreed on only 11.4% of the unique pairs they flagged.

That disagreement is the result: many apparent risks were artifacts of perspective compression, while the BEV method concentrated on physically close, rapidly closing vehicles.

Scatter plot comparing image-plane and bird’s-eye-view distance for flagged vehicle pairs — The top-left cluster is image-only: small pixel distance, but large real-world distance.

Flagged risk pairs per frame for both methods — Image-plane risk fires on 98.3% of frames; BEV risk on 39.4%.

Time-to-collision distribution for flagged vehicle pairs — BEV flags concentrate in the most urgent low-TTC range.

04 / ENGINEERING JUDGMENT

What held up

A minimum-track-age ablation reduced both event sets without changing the qualitative disagreement. A distance-threshold sweep showed that the result was not a quirk of one chosen cutoff.

What I’d improve

The model treats vehicles as point masses, assumes a flat ground plane, and uses a minimal four-point calibration. A production extension should use an overdetermined calibration, heading-aware geometry, uncertainty estimates, and benchmarked ground truth.

Threshold sensitivity curves for image-plane and geometry-aware methods — Sensitivity analysis confirms that perspective—not one threshold choice—drives the gap.

TEAM

Amaan Khan · Anirudha Kyathsandra · Dennis Binford · Mingyao Wang

CONTEXT

CS 543: Computer Vision · University of Illinois Urbana-Champaign

Source repository ↗Read the full report ↗More work ↗